LMCap: Few-shot Multilingual Image Captioning by Retrieval Augmented Language Model Prompting

TL;DR

LMCap introduces a retrieval-augmented, few-shot multilingual image captioning approach that leverages language models and retrieved captions, eliminating the need for large-scale multilingual caption datasets.

Contribution

It proposes a novel image-blind captioning method using retrieval and prompting, bypassing traditional training on multilingual caption data.

Findings

Competitive with fully-supervised models

No need for captioning training data

Effective across diverse geographic images

Abstract

Multilingual image captioning has recently been tackled by training with large-scale machine translated data, which is an expensive, noisy, and time-consuming process. Without requiring any multilingual caption data, we propose LMCap, an image-blind few-shot multilingual captioning model that works by prompting a language model with retrieved captions. Specifically, instead of following the standard encoder-decoder paradigm, given an image, LMCap first retrieves the captions of similar images using a multilingual CLIP encoder. These captions are then combined into a prompt for an XGLM decoder, in order to generate captions in the desired language. In other words, the generation model does not directly process the image, instead processing retrieved captions. Experiments on the XM3600 dataset of geographically diverse images show that our model is competitive with fully-supervised multilingual captioning models, without requiring any supervised training on any captioning data.